tht

About

Username
tht
Joined
Visits
195
Last Active
Roles
member
Points
8,028
Badges
1
Posts
6,023

Reactions

4.7KLike52Dislike663Informative

  • New iPad Pro rumored to debut with M4 chip

    thttht
    blastdoor said:
    nubus said:
    It doesn't make sense to ship any AI product before WWDC. There is no new iPadOS/AI to make use of that new AI power.
    Apple could release a new iPadOS at the same time they release a new iPad Pro, and that new version could be initially exclusive to the new hardware.

    Apple doesn’t initially need third party developer support for new AI stuff. It could initially be about improved Siri and other first party apps. 

    So, really, no need at all to wait for WWDC
    There is nothing unusual about the AI accelerators. It's just 8 or 16 bit integer SIMD, maybe 32 bit. Apple has been shipping these types of execution units for about 6 years now, and you could use the GPU, CPU to do it before that.

    With LLM chatbots breaking into mainstream consciousness, these SIMD units are now advertised as AI units in hope that people will buy new hardware. Since these LLM models will require a bit more performance than current SIMD and GPUs can muster, there is some optimism that AI features can drive an upgrade cycle. It all depends on AI providing features people would upgrade hardware for. I'm a bit skeptical. It's all part of the complex of compute hardware in the SoCs. There's a lot of transistors to spread around.

    A bit of roundabout way to say that the hardware is already here and Apple doesn't need to wait to ship SoCs with more NPUs. Furthermore, they made this decision about 3 years ago, at least. Who knows if these iPP models will come with the M4 SoC though. Kind of a big thing to keep secret. We should expect it to be M3 SoCs.
    Alex1Ncurrentinterest

  • Arizona TSMC facility continues to fight cultural battles, rising costs & logistical hurdl...

    thttht
    rob53 said:
    blastdoor said:
    I’d love to see intel regain the process lead and win apples foundry business. 
    Why? Intel was always slow to improve their products. They always run hot as well. 
    blastdoor is just saying that if Intel can regain the fabrication lead, he'd love it if Intel can fab Apple Silicon. He's not saying Apple will be using Intel designed processors. That is exactly what should happen. If Intel has the best fabs in the world, Apple should contract them to fab Apple Silicon, assuming the (cost x Watt)/mm2 all work out.

    Intel plans on catching up and passing TSMC. We all know what happens to plans, so probably not. But, Intel as a second source manufacturer of Apple Silicon isn't a bad idea either, or Samsung Semi. It would require Apple to sell more products, so long road to go on that.
    muthuk_vanalingamh2pblastdoor

  • New iPad Air & iPad Pro models are coming soon - what to expect

    thttht
    There have also been claims that Apple could offer a matte option for the display, providing an anti-glare coating to the tablet line for the first time.
    Ooh, a loaded iPad Pro 13" is sounding more expensive all the time. Would need to see what the matte option does to image clarity, and how well it cleans. Would finger oils just seep in between the cracks of the nanotexture surface, making it hard to wipe off the finger oils?
    gregoriusmAlex1Nwilliamlondon

  • When to expect every Mac to get the AI-based M4 processor

    thttht
    VictorMortimer said:
    jagrahax said:
    Half a terabyte is a lot of unified memory.  It might be worth checking this report against memory roadmaps from SK Hynix, Samsung, and Micron.  
    Not compared to the memory you can connect to an Intel or AMD chip, it isn't.  You already knew that, of course, because the REAL Mac Pro could use 1.5TB RAM, plus whatever video RAM was on your GPU boards.

    "Unified" memory just means it's sharing RAM between the CPU and GPU.  You know, the thing that used to be called "integrated graphics" that we used to acknowledge wasn't as good as a real GPU with its own RAM, but now that Apple does it, we pretend to fool ourselves?
    No, this is incorrect. 

    On-chip GPUs of the past partitioned off a section of system memory to serve as memory for the GPU. If a system has 8 GB of RAM, the on-chip GPU would partition off, say 1 GB, to serve as GPU memory, resulting in a decrease of system memory. To transfer GPU assets, programs would have to copy from system memory to the GPU memory partition. Some on-chip GPUs may have been limited on much 

    With Apple’s unified memory, there isn’t a partition and therefore there isn’t a transfer of assets to do. The GPU just accesses system memory wherever it needs to. Any processing unit in the SoC has uniform memory access to system memory.

    Modern x86 SoCs may have addressed this and is using unified memory architectures, I think? I don’t really know and need to go find out. 

    For discrete GPUs on a PCIe card, where transfers have to be made, there is limitation of how many MB could be transferred per copy/transfer command and introduces inefficiencies. Various companies are trying to institute protocols where a dGPU can access system memory directly and remove transfer size limits, like resizable bars or system addressable memory protocols, which enable a dGPU to address system right across the PCIe bus. 

    Regarding 192 GB to a possible 512 GB memory on an Apple SoC, the GPU will be able to access something like 80% of it. So an Apple GPU could have 150 to 400 GB of memory it can directly address.

    Most consumer dGPUs will be limited to something up to 32 GB. So it’s a natural advantage for certain workflows. Server GPUs have low hundreds of GB of memory, but not many people will have that on their desks. 
    baconstangAlex1NXed

  • Apple wants to make grooved keys to stop nasty finger oil transfer to MacBook Pro screens

    thttht
    Nothing but compromises. They can just increase the clearance between the top of the keys and the display, but nobody wants something like 1mm thicker device, right?
    9secondkox2williamlondon40domi